CN109508663A - A kind of pedestrian's recognition methods again based on multi-level supervision network - Google Patents
A kind of pedestrian's recognition methods again based on multi-level supervision network Download PDFInfo
- Publication number
- CN109508663A CN109508663A CN201811299473.0A CN201811299473A CN109508663A CN 109508663 A CN109508663 A CN 109508663A CN 201811299473 A CN201811299473 A CN 201811299473A CN 109508663 A CN109508663 A CN 109508663A
- Authority
- CN
- China
- Prior art keywords
- pedestrian
- network
- feature vector
- level supervision
- recognition methods
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to a kind of pedestrian's recognition methods again based on multi-level supervision network, this method extracts the feature of different semantic hierarchies to pedestrian image by a multi-level supervision network, and then realizes that pedestrian identifies again；The multi-level supervision network is including a multilayer depth convolutional neural networks as core network and multiple categorization modules as feature extraction sub-network；Pedestrian image is converted to the characteristic pattern of different semantic hierarchies by core network, each categorization module is separately converted to the feature vector with distinction by each layer characteristic pattern that supervised learning extracts core network, feature vector splicing shape on all levels is final feature vector, realizes that pedestrian identifies again based on the final feature vector.Compared with prior art, the present invention extracts the feature of pedestrian image difference semantic hierarchies, improves the distinction of feature, and the stability of training process is improved in the way of the supervised learning of half-separation type, network accuracy rate performance is improved, has many advantages, such as that heavy recognition accuracy is high.
Description
Technical field
The present invention relates to a kind of pedestrians again recognition methods, more particularly, to a kind of pedestrian's weight based on multi-level supervision network
Recognition methods.
Background technique
Pedestrian in video identify again be computer vision and artificial intelligence field an important subject.Its task
Target can summarize are as follows: the image (query image) of given one (or multiple) pedestrian to be searched needs having prison
All images of the pedestrian are searched out in control video image set (gallery images).Pedestrian identify again intelligent security guard,
The fields such as urban safety are of great immediate significance and are worth, and are a big hot spots of Recent study.
However, in reality scene, since video camera shooting angle, shooting distance, locating light environment etc. are different,
With a group traveling together, there are significant vision differences in different video.In addition to this, attitudes vibration caused by human motion, block
Phenomena such as further improve the difficulty of this task.Then, how pedestrian's image to be extracted under the influence of these factors and is had
The feature of high discrimination and be for identification this technical field problem.
Existing pedestrian's weight identification technology is commonly divided into 3 steps.Scheme firstly, preparing the largely pedestrian with identity label
Decent, as training set data library.Later, the depth convolutional neural networks of training set data training specific structure are utilized.Mind
Structure and training method through network generally determine the accuracy rate performance of pedestrian weight identifying system, are a most important steps.
Finally, using trained convolutional network to search library image zooming-out feature.When needing to identify pedestrian again, only need to utilize
Trained convolutional network extracts feature to pedestrian image to be checked (query image), by obtained feature vector and retrieval
The feature vector of library image compares cosine similarity or Euclidean distance one by one and is ranked up.The most similar several library images are
For the output result of pedestrian's weight identifying system.
Under the above technological frame, prior art focuses primarily on the structure design and instruction of depth convolutional neural networks
The mode of white silk designs.This kind of technology can substantially be divided into the pedestrian of two classes (1) based on provincial characteristics and identify again.(2) it is based on metric learning
Pedestrian identify again.
Pedestrian's weight identifying schemes based on provincial characteristics divide an image into multiple horizontal zones generally according to spatial position
Or net region.After region division, feature is extracted using depth convolutional neural networks to each block respectively.Feature mentions
The process taken may be summarized to be: original image (or block) is sent into convolutional neural networks, by multiple convolutional layers, batch normalization
The network units such as layer, nonlinear activation layer obtain the characteristic pattern containing high-level semantic feature.Later, by obtained characteristic pattern into
The global average pond of row, as the feature vector for representing this image block.Finally, the feature vector of all blocks is melted
It closes or merges, obtain the feature vector for representing this group traveling together.For example, document " Glad:global-local-alignment
descriptor for pedestrian retrieval”(Wei L,Zhang S,Yao H,et al.Proceedings of
The 2017ACM on Multimedia Conference.ACM, 2017:420-428) human body is drawn according to human body key point
It is divided into three head, the upper part of the body, lower part of the body regions, and respectively to these three extracted region features.Document " Beyond Part
Models:Person Retrieval with Refined Part Pooling”(Sun Y,Zheng L,Yang Y,et
Al.arXiv preprint arXiv:1711.09349,2017) pedestrian image level is divided into 6 blocks, and right respectively
Six extracted region features.
Pedestrian's weight identifying schemes based on metric learning are common usually using well-designed loss function training network
Metric learning loss function has control loss function, triple loss function, big distance s oftmax loss function etc..
There is also following disadvantages for the above-mentioned prior art:
1, during extracting feature, such methods only with depth convolutional network the last layer feature, to net
The utilization rate of network is not high.
2, such methods do not utilize multilayer semantic information caused by network.Although the characteristic pattern of network the last layer wraps
Containing stronger semantic information, but it also can be therefore easily lost image detail, and then cause the distinction of feature limited.
3, the pedestrian based on metric learning identifies that network is generally more difficult with training again.
Summary of the invention
It is an object of the present invention to overcome the above-mentioned drawbacks of the prior art and provide one kind based on multi-level prison
Superintend and direct pedestrian's recognition methods again of network.
An object of the present invention is to solve existing pedestrian weight identification technology to convolutional network middle layer characteristic use rate not
High problem improves the distinction and robustness of general characteristic.
The second object of the present invention is to improve the stability of network training process, and promote network accuracy rate performance.
The purpose of the present invention can be achieved through the following technical solutions:
A kind of pedestrian's recognition methods again based on multi-level supervision network, this method are based on a multi-level supervision network implementations
Pedestrian identifies that the multi-level supervision network includes a multilayer depth convolutional neural networks as core network and multiple points again
Pedestrian image is converted to the characteristic pattern of different semantic hierarchies as feature extraction sub-network, the core network by generic module, respectively
The categorization module is separately converted to the feature with distinction by each layer characteristic pattern that supervised learning extracts core network
Vector, the feature vector splicing shape on all levels is final feature vector, realizes that pedestrian knows again based on the final feature vector
Not.
Further, the multilayer depth convolutional neural networks are made of multiple residual error convolution modules, each residual error convolution mould
Block includes several convolutional layers, batch normalization layer and nonlinear activation layer.
Further, the input of multiple categorization modules respectively corresponds the defeated of multiple residual error convolution modules in core network
Out.
Further, each categorization module not shared parameter.
Further, the categorization module include the convolutional layer set gradually, it is batch normalization layer, nonlinear activation layer, complete
The average pond layer of office, dropout layers, the full articulamentum of dimension, batch normalization layer and softmax layers.
Further, the training process of the multi-level supervision network specifically includes:
1) by acquiring a collection of sample in pedestrian's database for collecting, input the multi-level supervision network carry out it is preceding to biography
It broadcasts；
2) intersect entropy loss according to the classification results of each categorization module and sample label calculating；
3) half-separation type backpropagation, the backpropagation are carried out to the multi-level supervision network based on intersection entropy loss
Cheng Zhong, only the last one corresponding categorization module of each convolutional neural networks module is completed to the reversed of entire multi-level supervision network
It propagates, remaining only carries out backpropagation to categorization module；
4) gradient descent algorithm and undated parameter are executed to network according to the gradient that backpropagation obtains, until network is received
It holds back.
Further, the final feature vector is spliced by the feature vector that each categorization module obtains.
Further, realize that pedestrian identifies again based on the final feature vector specifically:
By final feature vector and library image feature vector one by one compared with cosine similarity, with the highest preceding k of similarity
Image is as the query result identified again.
Further, the sample acquisition process in pedestrian's database specifically includes:
Pedestrian's video is obtained using different cameras by different spatial, is frame by pedestrian's video slicing, to every
Each of one frame image difference pedestrian marks an independent sample label, carries out resolution ratio scaling and reversion to every image
Processing forms pedestrian's database.
Compared with prior art, the present invention have with following the utility model has the advantages that
1, the present invention devises categorization module, realizes the depth convolutional network that can effectively extract different semantic hierarchies features
Structure merges multi-level merging features in identification, so as to improve the distinction and robustness of general characteristic, solves existing
The problem for having pedestrian's weight identification technology not high to convolutional network middle layer characteristic use rate, is more efficiently utilized computing resource.
2, the present invention devises a kind of effective multilayer supervision network training mode, the i.e. training method of half-separation type, has
Effect improves the stability of training process, and promotes network accuracy rate performance.
3, the inquiry phase identified again in pedestrian, the merging features Cheng Genggao for all levels that multiple categorization modules are obtained
The feature vector of dimension, the identification of lifting feature emphatically, and then significant improve the accuracy rate identified again.
4, the method for the present invention is suitble to pedestrian's weight identification mission under complex scene.
Detailed description of the invention
Fig. 1 is the structural schematic diagram of the multi-level supervision network of the present invention；
Fig. 2 is testing result schematic diagram of the invention.
Specific embodiment
The present invention is described in detail with specific embodiment below in conjunction with the accompanying drawings.The present embodiment is with technical solution of the present invention
Premised on implemented, the detailed implementation method and specific operation process are given, but protection scope of the present invention is not limited to
Following embodiments.
The invention proposes a kind of pedestrian's recognition methods again based on multi-level supervision network, in depth residual error network
(ResNet) on the basis of, using the categorization module (Classification Block) of multiple not shared parameters, in network
Supervised learning is carried out on different depth, and then the feature of different semantic hierarchies is extracted to pedestrian image.The overall structure of network is such as
Shown in Fig. 1.In the network training stage, the present invention uses a kind of supervised learning mode of half-separation type, improves training process
Stability, and improve network accuracy rate performance.In inquiry (query) stage that pedestrian identifies again, by the feature of all levels
It is spliced into more high-dimensional feature vector, emphatically the identification of lifting feature, and then significant improves the accuracy rate identified again.
The multi-level supervision network that the present invention uses include multilayer depth convolutional neural networks as core network with
Multiple categorization modules are turned residual error convolution module pedestrian image by a series of as feature extraction sub-network, the core network
It is changed to the characteristic pattern of different semantic hierarchies, each categorization module passes through each layer characteristic pattern that supervised learning extracts core network
It is separately converted to the feature vector with distinction, the feature vector splicing shape on all levels is final feature vector, is based on
The final feature vector realizes that pedestrian identifies again.The categorization module includes the convolutional layer set gradually, criticizes normalization layer, is non-thread
Property active coating, global average pond layer, dropout layer, the full articulamentum of dimension, criticize and normalize layer and softmax layers.Each classification
Module not shared parameter.
As shown in Figure 1, in the present embodiment, it is multi-level supervise network be based on existing ResNet50 (He K, Zhang X,
Ren S,et al.Deep residual learning for image recognition.Proceedings of the
IEEE conference on computer vision and pattern recognition.2016:770-778) it realizes.
ResNet50 is constituted with ResNet Block (residual error convolution module), and then extracts characteristic pattern to pedestrian image, and network is whole by 5
A stage composition, is referred to as Block1 to Block5.Each stage includes multiple residual error convolution modules, such as Block4 is by 6
A residual error convolution module composition, respectively Block4_1 to Block4_6；Block5 is made of 3 residual error convolution modules, respectively
For Block5_1 to Block5_3.These residual error convolution modules are by several convolutional layers, batch normalization layer, nonlinear activation layer structure
At.Traditional pedestrian identifies that network is usually only used to the feature of whole network the last layer (Block5_3) output again,
In the present embodiment, 9 categorization modules are provided with, are extracted Block4_1 to Block4_6, Block5_1 to Block5_3 totally 9
The feature of level, and be respectively fed to categorization module and exercise supervision study, to improve the distinction and robustness of feature.
In the present embodiment, the detailed process that categorization module carries out characteristic processing can be divided into:
(1) firstly, the triple channel pedestrian image that dimension is 256*128*3 enters network, by Block1, it is changed into 128*
The characteristic pattern of 64*64.
(2) characteristic pattern of 128*64*64 passes through Block2, is changed into the characteristic pattern of 64*32*256.
(3) characteristic pattern of 64*32*256 passes through Block3, is changed into the characteristic pattern of 32*16*512.
(4) characteristic pattern of 32*16*512 sequentially passes through Block4-1 to Block4-6, exports the feature of 6 16*8*1024
Figure.
(5) characteristic pattern of 16*8*1024 sequentially passes through Block5-1 to Block5-3, exports the feature of 3 16*8*2048
Figure.Different from original ResNet50, which removes the down-sampling operation of Block5.
The characteristic pattern of (6) 6 16*8*1024 is respectively fed to categorization module 1 to 6.Categorization module is 1*1* by core size
2048 convolutional layer, batch normalization layer, nonlinear activation layer, global average pond layer, dropout layers, 512 tie up full articulamentum,
Batch normalization layer, softmax layers form in order.It should be noted that the not shared parameters of categorization module 1 to 9.
(7) characteristic pattern of 16*8*1024 is converted into the characteristic pattern of 16*8*2048 through 1*1 convolutional layer, by the average pond of the overall situation
Changing layer becomes the feature vector of 1*1*2048, using full articulamentum dimensionality reduction, 512 dimensional feature vector of boil down to.
(8) 512 dimensional feature vectors are sent into softmax layers as the feature for representing pedestrian, complete the supervision of classification task
It practises.
(9) categorization module 7-9 is similar with categorization module 1 to 6, only difference is that Block5_1 to Block5_3 is exported
Characteristic pattern be 16*8*2048 tie up rather than 16*8*1024 tie up.
The present invention is trained multi-level supervision network using the training method of a half-separation type, improves training process
Stability, and promote network accuracy rate performance.Training process specifically includes:
Step 1: by acquiring a collection of sample in pedestrian's database for collecting, input the multi-level supervision network carry out it is preceding to
It propagates.
The collection of pedestrian's database includes: with preprocessing process
1) pedestrian's video needs to shoot in different spatial using different cameras, to guarantee the more of pedestrian image
Sample and otherness form the training set for possessing a large amount of different pedestrian's videos.
It 2) is that frame uses the mode or calculation manually marked for each frame image by video slicing after video collect
The mode of method automatic marking intercepts out the pedestrian in frame, saves as individual picture, and mark one to each different pedestrian
Independent sample label.
3) every picture is zoomed to the resolution sizes of 256*128, is needed with adapting to the input of subsequent convolutional neural networks
It asks, while being adapted to the physical size of human body, avoid anamorphose.
4) by treated, pedestrian's picture carries out horizontal 180 degree reversion one by one, to expand training data, is formed final
Pedestrian's database.
Step 2: intersecting entropy loss according to the classification results of each categorization module and sample label calculating.
Step 3: half-separation type backpropagation, the reversed biography being carried out to the multi-level supervision network based on entropy loss is intersected
During broadcasting, only the last one corresponding categorization module of each convolutional neural networks module is completed to entire multi-level supervision network
Backpropagation, remaining only carries out backpropagation to categorization module.
In the present embodiment, as shown in the dotted line of Fig. 1, for categorization module 1,2,3,4,5,7,8 backpropagation only to point
Generic module itself carries out, and the gradient of backpropagation is not back to the ResNet network of trunk, and dotted arrow indicates backpropagation ladder
Degree is truncated herein；For categorization module 6 and 9, categorization module itself and entire ResNet core network participate in reversely passing together
It broadcasts, completes the supervised learning of entire ResNet core network.
Step 4: gradient descent algorithm and undated parameter being executed to network according to the gradient that backpropagation obtains, until network
Convergence.
In the present embodiment, when carrying out pedestrian's weight identification mission using above-mentioned trained multi-level supervision network, one is given
Pedestrian image to be searched is opened, completes identification process according to following process:
(1) pedestrian image to be searched is zoomed to the resolution sizes of 256*128, with the good neural network of adaptation training
Input demand.
(2) pedestrian image is sent into trained multi-level supervision network and carries out propagated forward.Each categorization module output
The feature of one 512 dimension.
(3) feature vector that 9 categorization modules export is spliced, forms the feature vector of one 4608 dimension, represented
The final feature of pedestrian.
(4) by feature vector and the feature vector of library image of 4608 dimensions obtained in previous step one by one compared with cosine it is similar
Degree, and be ranked up according to similarity.The highest preceding k of similarity images are the query result of identification of attaching most importance to.K can be as needed
Setting.
Table 1 illustrates accuracy rate effect of the present invention on Duke-MTMC Reid [4] data set, it can be seen that the party
It is higher that method compares accuracy rate with some existing methods, and weight recognition effect is more preferable.In table, Rank@1, Rank@5, Rank@10 distinguish
1 before when indicating using CMC curve statistical, preceding 5, preceding 10 accuracy rate.
1 present invention of table is compared with prior art accuracy rate
Fig. 2 illustrates some effect of visualization of the present invention on Duke-MTMC Reid data set, effect of the present invention compared with
It is good.Query indicates that image to be searched, subsequent 10 picture are and the highest 10 library images of pedestrian's similarity.
The preferred embodiment of the present invention has been described in detail above.It should be appreciated that those skilled in the art without
It needs creative work according to the present invention can conceive and makes many modifications and variations.Therefore, all technologies in the art
Personnel are available by logical analysis, reasoning, or a limited experiment on the basis of existing technology under this invention's idea
Technical solution, all should be within the scope of protection determined by the claims.
Claims (9)
1. a kind of pedestrian's recognition methods again based on multi-level supervision network, which is characterized in that this method is based on a multi-level prison
It superintends and directs network implementations pedestrian to identify again, the multi-level supervision network includes a multilayer depth convolutional neural networks as backbone network
As feature extraction sub-network, pedestrian image is converted to different semantic hierarchies by the core network for network and multiple categorization modules
Characteristic pattern, each categorization module are separately converted to have and distinguish by each layer characteristic pattern that supervised learning extracts core network
Property feature vector, the feature vector splicing shape on all levels is final feature vector, is realized based on the final feature vector
Pedestrian identifies again.
2. pedestrian's recognition methods again according to claim 1 based on multi-level supervision network, which is characterized in that described more
Layer depth convolutional neural networks are made of multiple residual error convolution modules, and each residual error convolution module includes several convolutional layers, batch normalizing
Change layer and nonlinear activation layer.
3. pedestrian's recognition methods again according to claim 2 based on multi-level supervision network, which is characterized in that Duo Gesuo
The input for stating categorization module respectively corresponds the output of multiple residual error convolution modules in core network.
4. pedestrian's recognition methods again according to claim 1 based on multi-level supervision network, which is characterized in that each described
Categorization module not shared parameter.
5. pedestrian's recognition methods again according to claim 1 based on multi-level supervision network, which is characterized in that described point
Generic module include the convolutional layer set gradually, batch normalization layer, nonlinear activation layer, global average pond layer, dropout layers,
It ties up full articulamentum, criticize normalization layer and softmax layers.
6. pedestrian's recognition methods again according to claim 1 based on multi-level supervision network, which is characterized in that described more
The training process of level supervision network specifically includes:
1) by acquiring a collection of sample in pedestrian's database for collecting, the multi-level supervision network progress propagated forward is inputted；
2) intersect entropy loss according to the classification results of each categorization module and sample label calculating；
3) half-separation type backpropagation, the back-propagation process are carried out to the multi-level supervision network based on intersection entropy loss
In, only part classification module completes the backpropagation to entire core network, remaining only carries out backpropagation to categorization module；
4) gradient descent algorithm and undated parameter are executed to network according to the gradient that backpropagation obtains, until network convergence.
7. pedestrian's recognition methods again according to claim 1 based on multi-level supervision network, which is characterized in that it is described most
Whole feature vector is spliced by the feature vector that each categorization module obtains.
8. pedestrian's recognition methods again according to claim 1 based on multi-level supervision network, which is characterized in that be based on institute
It states final feature vector and realizes that pedestrian identifies again specifically:
By final feature vector and library image feature vector one by one compared with cosine similarity, with the highest preceding k of similarity images
As the query result identified again.
9. pedestrian's recognition methods again according to claim 6 based on multi-level supervision network, which is characterized in that the row
Sample acquisition process in personal data library specifically includes:
Pedestrian's video is obtained using different cameras by different spatial, is frame by pedestrian's video slicing, to each frame
Each of image difference pedestrian marks an independent sample label, carries out resolution ratio scaling and horizontal inversion to every image
Processing forms pedestrian's database.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811299473.0A CN109508663B (en) | 2018-10-31 | 2018-10-31 | Pedestrian re-identification method based on multi-level supervision network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811299473.0A CN109508663B (en) | 2018-10-31 | 2018-10-31 | Pedestrian re-identification method based on multi-level supervision network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109508663A true CN109508663A (en) | 2019-03-22 |
| CN109508663B CN109508663B (en) | 2021-07-13 |
Family
ID=65747474
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811299473.0A Active CN109508663B (en) | 2018-10-31 | 2018-10-31 | Pedestrian re-identification method based on multi-level supervision network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109508663B (en) |
Cited By (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110188635A (en) * | 2019-05-16 | 2019-08-30 | 南开大学 | A kind of plant pest recognition methods based on attention mechanism and multi-level convolution feature |
| CN110533184A (en) * | 2019-08-31 | 2019-12-03 | 南京人工智能高等研究院有限公司 | A kind of training method and device of network model |
| CN110781755A (en) * | 2019-09-28 | 2020-02-11 | 深圳中物智建科技有限公司 | Intelligent identification and metering method and system for construction steel bars |
| CN110969610A (en) * | 2019-12-03 | 2020-04-07 | 杭州天铂云科光电科技有限公司 | Power equipment infrared chart identification method and system based on deep learning |
| CN111291821A (en) * | 2020-02-20 | 2020-06-16 | 上海眼控科技股份有限公司 | Vehicle weight recognition method and device, computer equipment and storage medium |
| CN111340141A (en) * | 2020-04-20 | 2020-06-26 | 天津职业技术师范大学(中国职业培训指导教师进修中心) | Crop seedling and weed detection method and system based on deep learning |
| CN111414815A (en) * | 2020-03-04 | 2020-07-14 | 清华大学深圳国际研究生院 | Pedestrian re-identification network searching method and pedestrian re-identification method |
| CN112308000A (en) * | 2020-11-06 | 2021-02-02 | 安徽清新互联信息科技有限公司 | High-altitude parabolic detection method based on space-time information |
| CN112330743A (en) * | 2020-11-06 | 2021-02-05 | 安徽清新互联信息科技有限公司 | High-altitude parabolic detection method based on deep learning |
| CN112528756A (en) * | 2020-11-20 | 2021-03-19 | 华南理工大学 | Unsupervised pedestrian re-identification method based on heteromorphic image |
| CN112528059A (en) * | 2021-02-08 | 2021-03-19 | 南京理工大学 | Deep learning-based traffic target image retrieval method and device and readable medium |
| CN112733695A (en) * | 2021-01-04 | 2021-04-30 | 电子科技大学 | Unsupervised key frame selection method in pedestrian re-identification field |
| WO2021203801A1 (en) * | 2020-04-08 | 2021-10-14 | 苏州浪潮智能科技有限公司 | Person re-identification method and apparatus, electronic device, and storage medium |
| CN116912633A (en) * | 2023-09-12 | 2023-10-20 | 深圳须弥云图空间科技有限公司 | Training method and device for target tracking model |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106203318A (en) * | 2016-06-29 | 2016-12-07 | 浙江工商大学 | The camera network pedestrian recognition method merged based on multi-level depth characteristic |
| CN107103308A (en) * | 2017-05-24 | 2017-08-29 | 武汉大学 | A kind of pedestrian's recognition methods again learnt based on depth dimension from coarse to fine |
| CN107832672A (en) * | 2017-10-12 | 2018-03-23 | 北京航空航天大学 | A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information |
| US20180253596A1 (en) * | 2017-03-06 | 2018-09-06 | Conduent Business Services, Llc | System and method for person re-identification using overhead view images |
-
2018
- 2018-10-31 CN CN201811299473.0A patent/CN109508663B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106203318A (en) * | 2016-06-29 | 2016-12-07 | 浙江工商大学 | The camera network pedestrian recognition method merged based on multi-level depth characteristic |
| US20180253596A1 (en) * | 2017-03-06 | 2018-09-06 | Conduent Business Services, Llc | System and method for person re-identification using overhead view images |
| CN107103308A (en) * | 2017-05-24 | 2017-08-29 | 武汉大学 | A kind of pedestrian's recognition methods again learnt based on depth dimension from coarse to fine |
| CN107832672A (en) * | 2017-10-12 | 2018-03-23 | 北京航空航天大学 | A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information |
Non-Patent Citations (2)
| Title |
|---|
| FEI JIANG 等: "GRAPH REGULARIZED TENSOR SPARSE CODING FOR IMAGE REPRESENTATION", 《PROCEEDINGS OF THE IEEE INTERNATIONAL CONFERENCE ON MULTIMEDIA AND EXPO (ICME) 2017》 * |
| XIAOBIN CHANG等: "Multi-level Factorisation Net for Person Re-identification", 《2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
Cited By (21)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110188635B (en) * | 2019-05-16 | 2021-04-30 | 南开大学 | Plant disease and insect pest identification method based on attention mechanism and multi-level convolution characteristics |
| CN110188635A (en) * | 2019-05-16 | 2019-08-30 | 南开大学 | A kind of plant pest recognition methods based on attention mechanism and multi-level convolution feature |
| CN110533184A (en) * | 2019-08-31 | 2019-12-03 | 南京人工智能高等研究院有限公司 | A kind of training method and device of network model |
| CN110781755A (en) * | 2019-09-28 | 2020-02-11 | 深圳中物智建科技有限公司 | Intelligent identification and metering method and system for construction steel bars |
| CN110969610A (en) * | 2019-12-03 | 2020-04-07 | 杭州天铂云科光电科技有限公司 | Power equipment infrared chart identification method and system based on deep learning |
| CN111291821A (en) * | 2020-02-20 | 2020-06-16 | 上海眼控科技股份有限公司 | Vehicle weight recognition method and device, computer equipment and storage medium |
| CN111414815A (en) * | 2020-03-04 | 2020-07-14 | 清华大学深圳国际研究生院 | Pedestrian re-identification network searching method and pedestrian re-identification method |
| CN111414815B (en) * | 2020-03-04 | 2023-11-14 | 清华大学深圳国际研究生院 | Pedestrian re-recognition network searching method and pedestrian re-recognition method |
| WO2021203801A1 (en) * | 2020-04-08 | 2021-10-14 | 苏州浪潮智能科技有限公司 | Person re-identification method and apparatus, electronic device, and storage medium |
| CN111340141A (en) * | 2020-04-20 | 2020-06-26 | 天津职业技术师范大学(中国职业培训指导教师进修中心) | Crop seedling and weed detection method and system based on deep learning |
| CN112330743B (en) * | 2020-11-06 | 2023-03-10 | 安徽清新互联信息科技有限公司 | High-altitude parabolic detection method based on deep learning |
| CN112308000B (en) * | 2020-11-06 | 2023-03-07 | 安徽清新互联信息科技有限公司 | High-altitude parabolic detection method based on space-time information |
| CN112330743A (en) * | 2020-11-06 | 2021-02-05 | 安徽清新互联信息科技有限公司 | High-altitude parabolic detection method based on deep learning |
| CN112308000A (en) * | 2020-11-06 | 2021-02-02 | 安徽清新互联信息科技有限公司 | High-altitude parabolic detection method based on space-time information |
| CN112528756A (en) * | 2020-11-20 | 2021-03-19 | 华南理工大学 | Unsupervised pedestrian re-identification method based on heteromorphic image |
| CN112528756B (en) * | 2020-11-20 | 2023-06-23 | 华南理工大学 | Unsupervised pedestrian re-identification method based on different composition |
| CN112733695A (en) * | 2021-01-04 | 2021-04-30 | 电子科技大学 | Unsupervised key frame selection method in pedestrian re-identification field |
| CN112733695B (en) * | 2021-01-04 | 2023-04-25 | 电子科技大学 | Unsupervised keyframe selection method in pedestrian re-identification field |
| CN112528059A (en) * | 2021-02-08 | 2021-03-19 | 南京理工大学 | Deep learning-based traffic target image retrieval method and device and readable medium |
| CN116912633A (en) * | 2023-09-12 | 2023-10-20 | 深圳须弥云图空间科技有限公司 | Training method and device for target tracking model |
| CN116912633B (en) * | 2023-09-12 | 2024-01-05 | 深圳须弥云图空间科技有限公司 | Training method and device for target tracking model |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109508663B (en) | 2021-07-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109508663A (en) | A kind of pedestrian's recognition methods again based on multi-level supervision network | |
| CN110427813B (en) | Pedestrian re-recognition method of twin generation type countermeasure network based on gesture guidance pedestrian image generation | |
| CN112651262B (en) | Cross-modal pedestrian re-identification method based on self-adaptive pedestrian alignment | |
| CN101493887B (en) | Eyebrow image segmentation method based on semi-supervision learning and Hash index | |
| CN113963032A (en) | Twin network structure target tracking method fusing target re-identification | |
| CN109492528A (en) | A kind of recognition methods again of the pedestrian based on gaussian sum depth characteristic | |
| CN111723600B (en) | Pedestrian re-recognition feature descriptor based on multi-task learning | |
| Wang et al. | Global relation reasoning graph convolutional networks for human pose estimation | |
| El-Ghaish et al. | Human action recognition based on integrating body pose, part shape, and motion | |
| CN112434599A (en) | Pedestrian re-identification method based on random shielding recovery of noise channel | |
| CN111291612A (en) | Pedestrian re-identification method and device based on multi-person multi-camera tracking | |
| CN113221770A (en) | Cross-domain pedestrian re-identification method and system based on multi-feature hybrid learning | |
| CN113761995A (en) | Cross-mode pedestrian re-identification method based on double-transformation alignment and blocking | |
| CN114333062B (en) | Pedestrian re-recognition model training method based on heterogeneous dual networks and feature consistency | |
| CN117351518B (en) | Method and system for identifying unsupervised cross-modal pedestrian based on level difference | |
| CN112613474B (en) | Pedestrian re-identification method and device | |
| CN116453192A (en) | Self-attention shielding face recognition method based on blocking | |
| CN113869151B (en) | Cross-view gait recognition method and system based on feature fusion | |
| CN116935300A (en) | Gesture driving and dynamic alignment shielding pedestrian re-recognition method | |
| CN113435248A (en) | Mask face recognition base enhancement method, device, equipment and readable storage medium | |
| Tajik et al. | Video-based Person Re-Identification Using Attention Based on Separation Index | |
| Huang et al. | A Person Re-identification Method Fusing Bottleneck Transformer andRelation-aware Global Attention | |
| CN112712042B (en) | Pedestrian re-identification end-to-end network architecture with embedded key frame extraction | |
| Chang et al. | Improving Retrieval Efficiency of Person Re-Identification Based on Resnet50 | |
| CN112487232B (en) | Face retrieval method and related products |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| CB02 | Change of applicant information |
Address after: 200030 Dongchuan Road, Minhang District, Minhang District, Shanghai Applicant after: Shanghai Jiaotong University Address before: 200030 Huashan Road, Shanghai, No. 1954, No. Applicant before: Shanghai Jiaotong University |
|
| CB02 | Change of applicant information | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |